Known physics, learned from centuries of theoretical developments and numerous real-world applications, provides trustable input-output relation laws for behavior prediction of complex dynamic systems. Typically, such a prediction is obtained from a digital simulation, which will usually take much longer time than real time to run if the dynamic system is complex because of the required iterative numerical procedure for solving large sets of nonlinear differential-algebraic equations representing the dynamic system. We are currently developing a solution to this technology gap between the fast prediction need and inefficient physics-based simulation. The solution is to use a physics-based simulation model of a dynamic system to train a neural network-based deep learning model, such that the model can proximate the physics-based model for quick prediction of the dynamic behavior of the space system. This paper presents an early study result of this approach with a cart-pendulum system as an example. In the solution, we incorporate physics constraints to train a deep learning model of the dynamic system. The trained model shows promising results regarding the prediction speed and accuracy.